ci: build the AppImage on ubuntu:20.04, and re-budget the slow coop tests

[NonPolynomialTim]

Fixes build-linux, red on main since Debian 11 reached end of life, and re-budgets the coop tests that turned passes into reds on the clock.

build-linux

Two failures in sequence:

1. bullseye-security stopped being published on 2026-08-31 and its Release file expired 2026-09-07 21:13 UTC, so apt-get update failed outright. Before that it 404'd pool files its own index still advertised.
2. Pointing at archive.debian.org and dropping the dead suite fixed the fetch, but broke dependency resolution — the debian:11 image ships its packages at security versions and -dev packages need an exact = match of their runtime lib:

perl : Depends: perl-base (= 5.32.1-4+deb11u3) but 5.32.1-4+deb11u5 is to be installed
libstdc++-10-dev : Depends: libc6-dev (>= 2.30-1~) but it is not installable
E: Unable to correct problems, you have held broken packages.

Restoring the security suite is not available either — its pool is being dismantled unevenly across mirrors:

package	security.debian.org	deb.debian.org
libc6-dev +deb11u14	200	404
perl-base +deb11u5	404	200
libsepol1 3.1-1+deb11u1	404	404

No host serves a complete set.

Fix: ubuntu:20.04

glibc 2.31 with gcc 9.3 — the same floor debian:11 gave. Nobody loses support: Debian 11+, Ubuntu 20.04+, Mint 21 and RHEL/Rocky 9 (2.34) all still run the AppImage. All 19 build packages are present (patchelf and the SDL 1.2 -dev set in universe, enabled by default in the official image), and the project is C++17, which gcc 9.3 fully supports. It is also the image build-winxp already uses successfully in this pipeline, so its apt path is proven here.

debian:12 would fix apt too, but raises the floor to 2.36 and drops Ubuntu 22.04 LTS (2.35), Mint 21 and RHEL/Rocky 9 — a player support decision, not an apt workaround.

Applied to both ci-main.yml and ci-validate.yml.

Budgets (issue #169 roster)

test	budget	observed
test_coop_outcome_gaps	360 → 600	passed at 416.5s and 415.4s
test_parallel_loose_death	300 → 500	452.3s / 260.2s / 205.9s under load; 139–167s unloaded
test_parallel_heavy_death_repro	(none) → 400	180.9–241.2s; the only over-budget test with no exception

All three ran over during a window where unrelated tests were also at 1.5–2.5× their baselines — a slow/contended runner pool, not a behaviour change. outcome_gaps and heavy_death_repro were failing purely on the clock while passing.

heavy_death_repro stays quarantined (it is a repro tool, not a guard — exit 0 means the desync reproduced); the budget only keeps its KNOWN-FAIL line honest.

Not addressed here

test_sync_check (smoke bucket alarm) and test_parallel_soak (PRD-P2 drift tripwire after an alien side) are still red on main. Those are real drift detections in the parallel battlescape, same family as #168 / #178 / #179 / #182 — not timing, and not something this PR touches.